Why do passengers choose a specific car of a metro train during the morning peak hours?

Abstract

Crowding on metro trains is an important measure of passenger satisfaction and also provides a criterion for determining service frequency and the number of cars necessary for a train set. Particularly in metropolitan areas during morning peak hours, many studies have revealed a considerable difference in the crowding of specific cars on a single train. To accommodate the impact of this phenomenon in calculating metro capacity, a loading diversity factor has been adopted in many transportation studies. However, the underlying causes behind the uneven nature of carriage loading have rarely been examined in a systematic manner. In particular, there has been no trial to explain the nature of choice within a framework for individual passengers. Under the assumption that the uneven selection might stem from each passenger’s intrinsic preference for a specific car, the present study established a nested logit model to investigate the potential factors affecting the choice of a specific car on a train. Passengers were interviewed as they boarded from the platforms of line 7 of the Seoul Metro during the morning peak hours. Results show that the motivation to minimize the walking distance at destination stations turned out to be the most decisive in determining a passenger’s choice for a specific car of a train.

Full text

1
Why do passengers choose a specific car of a metro train during the morning
peak hours?
Hyunmi Kim
Graduate student
Department of Urban Engineering, Chung-Ang University, Seoul, Korea
221, Heukseok-dong, Dongjak-gu, Seoul 156-756, Korea
Sohee Kwon
Undergraduate student
Department of Urban Engineering, Chung-Ang University, Seoul, Korea
221, Heukseok-dong, Dongjak-gu, Seoul 156-756, Korea
emsjxnal@naver.com
Seung Kook Wu
Research Fellow
The Korea Transport Institute
315 Goyangdaero, Ilsanseo-gu, Goyang-si, Gyeonggi-do, Korea
email: wsk115@koti.re.kr
Keemin Sohn* Corresponding author
Associate Professor
Department of Urban Engineering, Chung-Ang University, Seoul, Korea
221, Heukseok-dong, Dongjak-gu, Seoul 156-756, Korea
kmsohn@cau.ac.kr

2
Abstract
Crowding on metro trains is an important measure of passenger satisfaction and also provides a
criterion for determining service frequency and the number of cars necessary for a train set.
Particularly in metropolitan areas during morning peak hours, many studies have revealed a
considerable difference in the crowding of specific cars on a single train. To accommodate the
impact of this phenomenon in calculating metro capacity, a loading diversity factor has been
adopted in many transportation studies. However, the underlying causes behind the uneven
nature of carriage loading have rarely been examined in a systematic manner. In particular, there
has been no trial to explain the nature of choice within a framework for individual passengers.
Under the assumption that the uneven selection might stem from each passenger’s intrinsic
preference for a specific car, the present study established a nested logit model to investigate the
potential factors affecting the choice of a specific car on a train. Passengers were interviewed as
they boarded from the platforms of line 7 of the Seoul Metro during the morning peak hours.
Results show that the motivation to minimize the walking distance at destination stations turned
out to be the most decisive in determining a passenger’s choice for a specific car of a train.
Keywords: Metro crowding; Loading diversity factor; Railway capacity; Train car choice; Nested
logit model; Latent variable

3
1. Introduction
Crowding within a train is a determinant of both the service level for passengers and the
supply level for operators. Thus, overcrowding in the morning peak hours has been recognized as
a serious problem for metro systems in urban areas (Hirsch and Thomson, 2011; Currie 2010; Hale
and Charles 2009; Qi et al., 2008). It is apparent that the crowding levels differ across individual
cars constituting a single train set. According to the transit capacity and quality of service manual
(TCQSM) (TCRP report 100, 2003), the passenger load imbalance between cars on individual trains
ranges from +61% to -33% with respect to the average passenger load per car in the Vancouver
SkyTrain, and fluctuates even more (from +156 to -89%) in Toronto’s Yonge Street subway. The
survey conducted for the present study showed that the loading difference varied from +118% to
-90% for the Seoul Metro line 7 during the morning peak hours. These phenomena commonly
imply that the practical capacity of rail transit could be overestimated without properly taking into
account the loading differences between the individual cars of a train. Vuchic (1981) suggested
the concept of a loading diversity factor to adjust the imbalance when calculating rail transit
capacity, which has been adopted in subsequent studies and in a series of transit capacity
manuals (TCRP report 100, 2003; TCRP report 13, 1996; Pudney and Wardrop, 2010; Jong et al.,
2011). According to the survey conducted by Pudney and Wardrop (2010), typical values for the
factor are 0.8 for heavy rail, 0.75 for light rail, and 0.6 for electric commuter trains. The latter
figure connotes that no less than 40% of capacity reduction is expected for metro services due to
the uneven passenger load across cars. The uneven utilization of cars not only decreases metro
capacity, but it also imparts great disadvantages in the maintenance of rolling stock.
Although the imbalance itself has often been used to exactly derive rail transit capacities,
little attention has been paid to determining the underlying causes behind uneven passenger
distribution across the individual cars of a train. Identifying the latent causes is expected to help
foment policies to distribute passengers more evenly across the individual cars of a train. For
example, there could be measures taken to accomplish a more even distribution by controlling
passengers’ behaviors and attitudes. Furthermore, it might be possible to streamline the
dispatching operation of a train. At least for a projected metro line, the infrastructure and train
design could be optimized. The goal of the present investigation was to find potential reasons for
the imbalance and, by controlling them, to get a more even distribution of passenger load across
cars.
It can be inferred that the unevenness of passenger loading can be attributed to the train
length, the platform layout, the service frequency, the entrance/exit location, and so forth.
However, there is no clear indication of the relative intensity of passenger preference for a specific
train car. Furthermore, to the best of our knowledge, individual characteristics and psychometric
propensities of metro passengers have never been investigated in association with the choice of a
specific car of a train. The present study provides new insight into the potential motives of

4
passengers to choose a specific car of a train within a hierarchical choice structure of individual
passengers. The direct objective of the present study was to identify possible reasons for the
unevenness of passenger loads across individual cars of a train, which can be utilized for
recommending measures to balance the passenger load more evenly.
The next section describes how variables affecting train car choice were selected, followed
by the data collection methodology. A conceptual framework and details of the choice structure is
described based on the nested logit model in the fourth section. The model results are discussed
in the fourth section. The last section draws conclusions and suggests several measures to
promote the even distribution of metro passenger loading across the cars of a train. Further
research necessary for advancing the results of the present study is also proposed in the last
section.

5
2. Determining potential variables
Potential variables that were expected to affect the passenger preference for a specific car of
a train were determined based on our pilot survey and a review of the previous literature dealing
with conventional mode or route-choice analyses (Sohn and Yun, 2009; Lee et al., 2012; Ben-Elia
et al., 2008; Papinski et al., 2009). The chosen variables can be categorized into four groups:
individual-specific characteristics, trip-related variables, physical environment, and attitudinal or
behavioral propensities.
2.1. Individual-specific characteristics
Conventionally, the individual-specific variables in Table 1 have been regarded as key factors
affecting a traveler’s decision-making. The background of adopting some unique variables in this
category needs to be addressed. Passengers with physical handicaps were expected to find a less
crowded car and to try to reduce their walking distance. Obesity might have passengers walking
shorter distances or finding a less crowded car.
2.2 Trip-related variables
Variables related to a respondent’s current trip encompassed trip purpose, trip frequency,
travel times, prior travel experience, number of transfers, awareness of station layout, and transfer
station (see Table 2). It should be noted that the variables were defined for each respondent with
respect to his/her current trip. Two variables regarding transfers were included to test if a transfer
passenger had a different preference on choosing a car from non-transfer passengers. While the
transfer station variable indicated whether a passenger transferred at stations of Line 7, the
number of transfers included transfers between other lines within a subject’s itinerary. There was
an expectation that passengers in a party would take on different attitudes when choosing a car
of metro trains. Passengers laden with luggage were expected to find a less-crowded car and to
try to reduce their walking distance. The variable did not cover handbags under the assumption
that women would feel little burden from them. Women in high heels could also be more
sensitive to the walking distance. The passenger choice for a car was assumed to vary according
to the characteristics of their trip.
2.3 Physical environment
The physical environment around a platform contains entrances/exits, transfer gates, elevators,
and escalators. A respondent’s accessibility to each facility was quantified as the walking distance
between the facility and his/her selected car. The distances were defined only for each
respondent’s origin and destination stations. These variables were hypothesized to affect the
preference of passengers on a specific car of a train. As another physical environment, the
crowding within individual cars was expected to be a key factor in the choice of a specific car.

6
Unfortunately, there was no direct way to measure it. Instead, the axle load was available and
employed as a surrogate for crowding, which was measured before loading and unloading
passengers when a train stopped at each station.

7
Table 1. Individual-specific variables and the corresponding data description
Variable
Total
Maximize walking
distance at destination
Maximize walking
distance at origin
Maximize comfort
Choose a car
unintentionally
Mean
Standard
deviation
Mean
Standard
deviation
Mean
Standard
deviation
Mean
Standard
deviation
Mean
Standard
deviation
Age
37.050
15.818
36.204
15.050
30.077
12.367
36.000
10.583
40.000
17.317
Gender(Male: 1; Female: 0)
0.494
0.501
0.435
0.497
0.500
0.510
0.400
0.548
0.593
0.493
Marital status
(Married: 1; Unmarried: 0)
0.497
0.501
0.503
0.501
0.654
0.485
0.000
0.000
0.475
0.501
Number of children
0.453
0.516
0.435
0.518
0.269
0.452
0.600
0.548
0.517
0.519
Income (More than 30 million won:
1; Less than 30 million won: 0)
0.485
0.501
0.518
0.501
0.462
0.508
1.000
0.000
0.415
0.495
Occupation 1
(Student: 1; Otherwise: 0)
0.338
0.474
0.372
0.610
0.500
0.510
0.000
0.000
0.314
0.466
Occupation 2(White color: 1;
Otherwise: 0)
0.374
0.484
0.414
0.494
0.346
0.485
0.600
0.548
0.305
0.462
Education (College-graduated: 1;
Otherwise: 0)
0.426
0.495
0.440
0.498
0.385
0.496
0.800
0.447
0.398
0.492
Car ownership
0.379
0.559
0.346
0.509
0.385
0.637
0.800
0.447
0.415
0.618
Dwelling type (Apartment: 1;
Otherwise:0)
0.565
0.497
2.487
0.631
2.692
0.618
2.600
0.894
2.525
0.713
Obesity
(Obesity: 1; Otherwise: 0)
0.368
0.483
0.387
0.488
0.231
0.430
1.000
0.000
0.339
0.475
Physical handicap(Inconvenience: 1,
Otherwise: 0)
0.035
0.185
0.042
0.201
0.000
0.000
0.200
0.447
0.025
0.158
Sample size
340
181
43
36
80

8
Table 2. Trip-related variables and the corresponding data description
Variable
Total
Maximize walking
distance at destination
Maximize walking
distance at origin
Maximize comfort
Choose a car
unintentionally
Mean
Standard
deviation
Mean
Standard
deviation
Mean
Standard
deviation
Mean
Standard
deviation
Mean
Standard
deviation
Trip purpose (Discretionary: 1,
Compulsory: 0)
0.253
0.435
0.168
0.832
0.231
0.430
0.000
0.000
0.593
0.493
Trip frequency
4.532
2.417
4.890
4.890
4.808
2.117
6.800
4.087
3.797
2.590
Travel times
38.985
19.454
39.770
39.770
34.731
12.951
43.000
12.042
38.483
21.407
Prior travel experience
0.850
0.358
0.916
0.278
0.808
0.402
1.000
0.000
0.746
0.437
Number of transfers
0.685
0.602
0.675
0.675
0.692
0.471
1.000
0.000
0.492
0.502
Awareness of station layout
0.850
0.365
0.942
0.234
0.962
0.196
0.800
0.447
0.678
0.487
Transfer station
0.617
0.487
0.675
0.675
0.692
0.471
1.000
0.000
0.492
0.502
Passenger laden with luggage(with
luggage: 1; Otherwise: 0)
0.324
0.469
0.335
0.473
0.269
0.452
0.000
0.000
0.331
0.472
Uncomfortable shoes (high
heels)(Uncomfortable shoes: 1;
Otherwise: 0)
0.3
0.459
0.304
0.461
0.346
0.485
0.400
0.548
0.280
0.451
Number of accompanying persons
0.088
0.323
0.068
0.309
0.154
0.368
0.000
0.000
0.110
0.341
Sample size
340
181
43
36
80

9
2.4 Attitudinal or behavioral propensities
Attitudinal or behavioral propensities were widely adopted to enhance the explanatory power
of choice models in the field of mode and route-choice studies (Prato et al., 2012; Sohn and Yun,
2009; Johansson et al. 2006; Morikawa et al. 2002; Ben-Akiva et al. 2002, 2006). Among indicators
adopted in the literature, the present study selected 20 indicators in advance (see Table 3), each
of which was a simple question that was expected to be associated with passenger behaviors and
attitudes toward the choice for a specific car of a train. Indicators of the first factor in Table 3
were adopted with the intention of identifying how different a passenger’s car choice behavior is
depending on whether he/she is a heavy user of portable electronic devices. A commercial agency
is currently providing metro passengers in Seoul with real-time information about the location of
entrances/exits and escalators/elevators for their current trip, which would encourage passengers
to use a portable electronic device in choosing their own specific car. Indicators of the second
factor reflected the ability to remember, with the expectation that those who have a good
memory might have advantages in the use of dispersed station facilities. Indicators of the third
factor were included to identify how the health condition of a passenger affects his/her car choice.
Indicators of the fourth factor were associated with a passenger’s sense of time. The remaining
indicators were selected since they were expected to be related to a passenger’s sociability,
activeness, and patience, which also might affect car choice.
Table 3. Psychometric propensities and their indicators
Variable
Indicators
1
Preference for
electronic devices
I am accustomed to searching for something with a smartphone; I am an early adapter
for IT products; I usually get information from newspapers rather than the Internet*; I
am a morning person*.
2
Mnemonic ability
I seldom forget the road I have traveled on; I can easily find things in the dark; I only
use the road that I know*.
3
Health condition
I can climb two stories of a building without a break; I take medicine for chronic
diseases like diabetes or hypertension*; I walk on escalators.
4
Punctuality
I always arrive at the expected time; I can’t understand those who are late for work
frequently; I can remember the birthdays of my friends.
5
Antisociality
I feel comfortable when I solely shop.
6
Impatience
I am upset if someone overtakes me while driving; I can’t wait in a long queue for
dinner or lunch; I feel irritated when I’m in a crowded situation.
7
Active propensity
I enjoy exercise on a regular basis; I often meet someone on business for dinner.
Only indicators that had a factor loading value larger than 0.5 are shown.
* : a negative factor loading.

10
3. Survey and data acquisition
3.1. Overview of survey
The test-bed of the present study was line 7 of the Seoul metro. Line 7 is 46.9 km long with 41
stations and is operated at an average speed of 32.3 (km/h). The line connects from the north-
east to the south-west via the Gangnam area, the second biggest employment center in the Seoul
metropolitan area. The line dispatches eight-car trains every 5 minutes during peak hours. It takes
87 minutes to run the line in one direction. Out of 41 stations of the line, 11 are transfer stations
where passengers can transfer to or from other lines. The morning peak hours of line 7 range
from 7 AM to 9 AM, which was when the survey was conducted. The physical layout of each
station was examined to identify the location of facilities. Line7 has four types of station layouts.
Each type shares common locations of facilities such as entrances/exits, elevators/escalators, and
transfer gates.
We randomly selected respondents, so that the entire sample could be distributed
proportional to the number of boarding passengers at each station. The sample size was 340,
after eliminating data associated with incomplete or contradictory replies. Because of the lack of
surveyors, it took four weeks to collect the sample. For weekdays of the period, three surveyors
chose their own convenient days and conducted the survey during the morning peak hours. They
were instructed to randomly choose a subject on a platform, considering the distribution of
waiting passengers across each car’s stop position. However, this created a possible bias in the
sample, since surveyors had to depend on their own visual inspection to find the distribution of
passengers. Another bias was an abnormal concentration of subjects in a particular station, which
was due to the convenience of the station location that happened to be closest to our laboratory.
The results were, however, certainly free from bias since the study focused only on the car-choice
behavior of metro riders, which was unlikely to differ from the point of boarding.
3.2. Data for dependent variables
The most important thing in the survey was to collect data for the dependent variables of
the model. The dependent variable was set as the respondents’ responses to the following two
questions. The first question asked passengers, who were waiting for boarding in the platform,
whether or not they chose a specific car intentionally. Those who answered “yes” were asked a
subsequent question concerning their motivation for the choice. Three main options for the
motivation were confirmed by a week-long pilot survey in advance. Although a finer
categorization of motivations could have been done, these three options were finally adopted
because, in the context of statistical modeling, the parsimony was important and there was
actually an ignorable number of respondents who had a distinct motivation that should fall in a
category other than one of the three motivations. As a result, 76.6% of the respondents reported
choosing a specific car intentionally. Among them, 69.7% stated that their motivation was to

11
minimize the walking distance to exit when they disembarked at a destination station, 16.6%
reported that they sought to minimize the distance from the entrance when they boarded at an
origin station, and the remaining 13.5% reported that they wanted to pursue comfort while
traveling. There were also minor groups who had other motivations in the main survey, but they
were negligible and eliminated from the sample.
3.3. Data for independent variables
3.3.1. Individual-specific data
Data descriptions for the independent variables that fell on the individual-specific
characteristics are shown in Table 1. Surveying the variables did not depend on a self-answering
questionnaire to avoid a potential bias. Rather, the first ten variables in Table 1 were collected
directly from the interview, while the remaining two variables were observed by the sur veyor.
Despite the possibility that this would cause some bias associated with surveyors, there was no
robust way to survey the variables without a privacy infringement. Surveyors were instructed to
see the walking pace of subjects when they entering the platform and to classify them to have
obesity if their pace was much slower than the average pace at which most other people walk. A
physical handicap was defined as the disability to walk. A person who appeared to be lame or
used a wheelchair or crutches was classified as handicapped.
3.3.2. Trip-related data
Data for variables related to a respondent’s current trip was also obtained by interview except
for the variable of uncomfortable shoes. To collect data for this variable, surveyors used physical
observation to approximate the height of the shoes, and shoes that seem higher than 7cm fell
into the high-heel category. Each surveyor asked respondents their trip purpose, trip frequency,
travel times, prior travel experience, number of transfers, awareness of station layout, and transfer
station. Five categories of trip purpose were predetermined through the pilot study, so that they
could be used in the subsequent modeling. Actually, the five trip purposes were reclassified into
two categories of trip purpose and were taken into account in the modeling: compulsory or
discretionary trips. Compulsory trips encompassed commuting, work, and school trips, while
discretionary trips were for shopping, leisure, and visiting purposes (Otuzar and Willumsen, 2004).
The awareness of station layout was measured with a five-point Likert scale. For every metro line,
there were transfer stations where passengers could transfer to or from other lines. Since there is
no cross-platform transfer in line 7, every transfer station had one or two gates linked to other
lines, which was not quite different from exits/entrances. The dummy variable for transfer station
was set at 1 only for respondents who transferred. Origin and destination stations for a
respondent’s trip were confined to line 7, which, however, does not mean transferred passengers
were excluded. The confinement was only for brevity in modeling. Four types of metro trips were

12
identified in the survey. The first type encompassed trips that used only Line 7. In this case, there
was no difficulty in determining their origin and destination stations. The second type
corresponded to trips that started from other lines, transferred at a station of Line 7, and
terminated at another station of Line 7. In these cases, the transfer station was regarded as an
origin station. The third type included trips that started at a station of line 7 and transferred from
another station of line 7 to other lines. In these cases, the transfer station was regarded as a
destination station. The last category included trips that started and terminated at stations of
other lines. In these cases, both transfer stations were origin and destination stations, respectively.
In addition, the trip frequency had to be distinguished from the travel experience. The former was
applied for respondents with compulsory trips on a regular basis, while the latter was for
discretionary trips.
3.3.3. Data for the physical environment
Regarding the collection of variables indicating a respondent’s physical environment, all
physical aspects of each station had been examined in advance through the pilot survey, so that
these, afterwards, could be matched with subjects. To constitute the physical environment
variables, surveyors examined distances from a respondent’s selected car to each facility in his/her
origin and destination stations. Such distances were measured in the laboratory after finishing the
field survey. For convenience, rather than using the metric system, the distance was measured by
counting the number of cars along a platform between a facility and the stopping position of a
respondent’s selected car. The distance ranged from 1 to 8 since line 7 of the Seoul Metro, which
was a test bed for the present investigation, dispatched eight-car trains in the morning peak
hours. The accessibility variable, which was actually used in the modeling, took on the inverse
value of the distance variable, so that the variable for the longest distance was set at 1, the
variable for the shortest distance at 8, and if the corresponding facility was unavailable or
unnecessary for a respondent, the variable was set at zero. For example, a passenger who
transferred at a station to other lines tried not to reduce the distance to the exit but, rather, to
reduce the distance to the transfer gate. To accommodate this in the modeling, such a respondent
was endowed with zero accessibility to the exit.
Unfortunately, data for axle loads measured in the peak hours was available only for five
weekdays in a row that fell on the one-month sur vey period. Fig. 1 shows the average axle load
distribution for the available data across individual cars. As expected, a considerable difference
across individual cars was confirmed. In particular, the discrepancy was exaggerated in consecutive
stations with the same layout where the overall passenger load was high (stations 28 to 36 for
south-westbound trains and stations 9 to 16 for north-eastbound trains). For convenience, rather
than directly using the weight, the axle load variable took on a rank in the weight across
individual cars, which ranged from 8 to 1 in a decreasing order of axle load. That is, the value of 1
was assigned to the most crowded (=heaviest) car of a single train, and the value of 8 to the least

13
crowded (=lightest car). To constitute the axle load variable, surveyors assigned each respondent a
predetermined rank corresponding to his/her chosen car.
(a) Average axle loads for south-westbound trains
(b) Average axle loads for north-eastbound trains
Figure 1. Axle load distribution across the individual cars of line 7 in the morning rush hours
3.3.4. Data for latent variables
Indicators of passengers’ psychometric propensities were surveyed through a self-answering
questionnaire. For convenience, available answers to the questions were confined to a five-point
Likert scale that ranged from “strongly negative” to “strongly positive.” The scale included a
neutral option but excluded an unsure option. Respondents had to select one of the five options
for each indicator shown in Table 3. To reduce the number of variables, a principal component
analysis (PCA) was applied to the indicators. The PCA was useful in reducing many indicators into
a tractable number of factors (=variables). By convention, each factor dictated its main indicators
if the absolute value of a factor loading exceeded 0.5. Table 3 shows the resultant 7 variables and
their respective indicators after completing the PCA. Each variable was titled based on our
judgment after examining the corresponding indicators
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41
Axle load(kg)
Station number
car 1
car 2
car 3
car 4
car 5
car 6
car 7
car 8
standrad
deviation
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41
Axle load(kg)
Station number
car 1
car 2
car 3
car 4
car 5
car 6
car 7
car 8
standrad
deviation

14
4. Modeling and conceptual framework
Two hierarchical structures were hypothesized to address the passenger preference for a specific
car of a train. The first structure had a three-level hierarchy and the second had a two-level
hierarchy (see Fig. 2). As for the three-level structure, the top level accommodated the passenger
decision about whether to intentionally choose a specific car. At the middle level, the motivation
behind the car choice was classified into either minimizing walking distance or maximizing
comfort. At the bottom level, the point at which a passenger wanted to minimize his/her walking
distance was separated into his/her origin or destination stations. For the other two-level structure,
the middle and bottom levels of the three-level structure were incorporated as a nest. Actually,
after estimating two nested logit models, the three-level structure was collapsed to a two-level
structure because the inclusive value (IV) parameter for the bottom nest of the three-level model
did not significantly differ from unity.
(a) Three-level choice hierarchy
(a) Two-level choice hierarchy
Figure 2. Hypothesized choice structures
Prior to modeling the two-level choice structure, a conceptual framework was set up, which
reflected potential relationships between the selected variables and each end option in the
structure, as shown in Fig. 3. Individual-specific variables were expected to affect the upper-level
decision as to whether or not a passenger chose a specific car intentionally. Physical environments
were assumed to be associated with the lower-level decision on where a passenger would
minimize his/her walking distance given that he/she chose a specific car intentionally. Trip-related
and psychometric variables were hypothesized to have an influence on decisions at both levels. As
a whole, arrows that are augmented but toned down with a mild color in Fig. 3 indicate our own
Choose a car
unintentionally
Choose a car
intentionally
Distance at
origin
Minimize walking
distance Maximize comfort
Distance at
destination
Choose a car
unintentionally
Choose a car
intentionally
Minimize walking
distance at origin Maximize comfort
Minimize walking
distance at
destination

15
conceptual framework. However, numerous sub-combinations could be made with respect to the
relationships between individual variables and each end option. Slender arrows in Fig. 3 show the
final model specification after testing as many plausible models as possible within the conceptual
framework. As the comment attached to the figure indicates, dotted lines indicate a negative
influence on the linked motivation. That is, a dotted arrow from a variable to its motivation
corresponds to the inverse relationship between them.
The final model specification was statistically significant with an acceptable goodness-of-fit.
The evaluation of a nested logit model was three-fold (Ben-Akiva and Lerman, 1985). First, the
model structure can be validated based on the estimated coefficient for an IV. To rationalize a
model structure, the coefficient should differ from unity with a statistical significance at the 0.05
level. The value for the second-level nest of the present model was 0.389 at a statistical
significance level of 0.032. Second, the overall goodness-of-fit can be measured by
2
ρ
or
2
ρ
.
The former reflects the contribution of information (=modeling) with respect to the log-likelihood
without information. The latter is developed to adjust the impact of the number of coefficients to
be estimated. The indices for the present model were 0.287 and 0.273, respectively, which did not
deviate from other choice studies (Lee et al., 2012; Sohn and Yun, 2008). Last, the statistical
significance of influential variables determined the usability of the model. The results from the
present model will be discussed in the next section.
Dotted lines indicate their negative influence on the linked motivation.
Figure 3. Conceptual framework
Choose a car intentionally
Lower model
Minimize walking distance at
origin
Maximize comfort
Minimize walking distance at
destination
Age
Gender
Marital
status
Number of children
Income
Occupation
Education
Physical
handicap
Passenger laden with
luggage
Number of accompanying
persons
Uncomfortable
shoes
Obesity
Car ownership
Dwelling
type
Preference for electronic
devices
Early
-bird propensity
Mnemonic ability
Health condition
Punctuality
Independence
Impatience
Active propensity
Individual-specific variables Psychometric variables
Accessibility
entrance at
origin
Accessibility from transfer
gate at origin
Accessibility
from elevator
at origin
Accessibility
from escalator
at origin
Accessibility
to exit at
destination
Accessibility
to transfer
gate
at destination
Accessibility
to elevator at
destination
Accessibility
to escalator at
destination
Axle loads
Physical environment
variables
Upper model
Trip-related variables
Trip purpose
Number of transfers
if a
respondent transfers
Transfer station
Trip frequency
Travel time
Awareness
of station layout
Prior
travel experience
Choose a car unintentionally

16
5. Results and discussion
On the whole, the hypothesized choice structure of the two-level hierarchy turned out to be
sufficient to address passenger preferences for a specific car of a metro train, since the parameter
for inclusive value (IV) was sufficiently different from unity and statistically significant at the 0.05
level. The estimation results from the final model specification offered useful results to identify
potential motivations behind the passenger car choice in the morning peak hours. The estimation
results are shown in Table 4 and the impact of each variable on each level of decision will be
discussed as follows.
5.1. Results from the upper model
White collar young females with a compulsory trip purpose were more likely to choose a
specific car of a train during the morning peak hours, which was also consistent with the
descriptive statistics in Tables 1 and 2. This result also support the notion that young females
tended to avoid physical contact with strangers in a crowded car, which is consistent with the
results from a focus group interview conducted by Hirsch and Thompson (2011), which found that
young female passengers in particular regarded crowded train cars as opportunities for men to
surreptitiously take advantage of them. The proclivity that women are averse to a crowded metro
train was also confirmed in an effort by two Asian cities to separate women from men during the
peak hours of metro operation. The Seoul metropolitan government has introduced metro trains
with a specific car reserved only for women. Tokyo metro is now successfully operating such trains.
This policy was indirect evidence showing that women feel some kind of fear when they are with
strangers in a crowded train. As shown in Table 4, commuters with a compulsory trip purpose
tried to minimize their walking distance by selecting a specific car of a train. Among the trip-
related variables, a passenger’s awareness of the layout of his/her origin and destination stations
proved to expedite the preference for a specific car of a train. Unlike prior expectations, the
influence of a passenger‘s psychometric propensities was minimal when addressing potential
causes behind his/her car choice. Only a coefficient for the variable of impatience among
psychometric variables took on an intuitively accountable sign for the upper model. The
coefficient was, however, not backed up by statistical significance. A possible reason for the
statistical insignificance of the variable, as well as other latent variables, might be due to the
sequential estimation methodology, in which PCA was conducted to derive the latent variables
independently of the choice model.
5.2. Results from the lower model
The lower-level decision identified the direct motivation behind passenger car choice. Physical
environments were a key determinant for passenger car choice based on minimizing the walking
distance. The accessibility to entrances/exits was positively associated with the car choice that

17
depended on minimizing walking distance, even though the variable of the origin station had an
insufficient statistical significance that slightly exceeded the marginal value (=0.1). It was
interesting that at origin stations a transfer passenger’s accessibility from a transfer gate to his/her
boarding position had nothing to do with his/her car choice to minimize walking distance, while
at destination stations a transfer passenger’s accessibility from his/her alighting position to a
transfer gate was positively associated with the car choice to minimize walking distance. The
insignificant transfer accessibility at origin stations reflects the notion that the connection area
between a platform and a transfer gate was so crowded in the morning peak hours that
passengers from the transfer gate couldn’t board a car stopping close to the area. The
accessibility from escalators at origin stations was found to be positively associated with the
motivation of car choice to minimize the walking distance. The reason that accessibility to
escalators was insignificant at destination stations might also be due to the congestion around
escalators when passengers disembark after the arrival of a train. Consequently, passengers
tended to place a high priority on the walking distance at their destination station, regardless of
whether they transferred to other lines or terminated their trip there. In addition, the walking
accessibility was more important for passenger car choice at destination stations than at origin
stations, because at destination stations the variables were more statistically significant.
Among latent propensities and trip-related variables, only the activity was connected to
passenger car choice at origin stations at the marginal significance level, but the rationale for this
result eluded us. A latent variable and two trip-related variables were statistically significant for car
choices based on minimizing the walking distance at destination stations, and thus could be
additional grounds for separating the motivation of car choice into that at origin stations and that
at destination stations. The mnemonic ability turned out to be positively linked to the passenger
choice of a car to minimize the walking distance at destination stations. As expected, the trip
frequency positively affected the passenger car choice to minimize the walking distance in
destination stations. Passengers who had good memory and made a compulsory trip on a regular
basis should be more familiar with the layout of their destination station and, thus, a good
memory was more advantageous in minimizing the walking distance at destination stations. The
travel experiences proved to be positively associated with the passenger car choice based on the
minimization of walking distance at destination stations. This implies that, for discretionary trips,
the prior travel experience could help passengers find the shortest walking distance at his/her
destination station.
Unfortunately, estimated coefficients from the motivation of car choice to maximize comfort
were either marginally or negligibly significant. The model estimation that car choice was based
on comfort was expected to depend largely on travel times. That is, passengers with a long trip
were assumed to be more likely to select a less-crowded car. The estimated coefficient for the
travel time variable managed to match the expectations at the marginal significance level. A
serious problem occurred when dealing with the surrogate variable for crowding. According to the

18
sign of the estimated coefficient, the crowding rank variable derived from axle loads turned out to
expedite car choice depending on maximizing comfort, but the result had little statistical
significance. The reason for yielding an insignificant coefficient for the variable was addressable by
the possibility that the axle load data might have lost its accuracy while going through the
averaging process. The axle load data during the morning rush hours were available only for 5
days in a row, which were included in the entire 4-week survey period. If the exact crowded level
within an actual car chosen by a respondent had been available, the coefficient would have been
statistically significant.
Table 4. Estimation results of the two-level nested logit model
Variable
Coefficient
t-value
p-value
Choose a car unintentionally
Age
0.025***
2.568
0.010
Gender (Male: 1, Female: 0)
1.143***
3.167
0.002
Occupation (White collar: 1, Blue collar: 0)
-0.658**
-1.737
0.082
Trip purpose (Discretionary: 1, Compulsory: 0)
0.692**
1.695
0.090
Awareness of station layout
-1.985***
-4.663
0.000
Impatience
-0.193
-1.149
0.251
Minimize walking distance at origin
Active propensity
0.352**
1.735
0.083
Accessibility from entrance at origin
0.112
1.493
0.136
Accessibility from escalator at origin
0.110**
1.792
0.073
Minimize walking distance at destination
Mnemonic ability
0.323***
2.062
0.039
Accessibility to exit at destination
0.177***
2.577
0.010
Accessibility to transfer gate at destination
0.146***
2.605
0.009
Trip frequency
0.165***
2.244
0.025
Prior travel experience
0.586***
1.944
0.052
Maximize comfort
Travel time
0.015**
1.617
0.104
Rank in axle load
0.056
0.769
0.442
Nesting parameter
Inclusive value
0.389***
2.15
0.032
Goodness-of-fit
Maximum log likelihood
-294.252
Restricted log likelihood
-457.041
2
ρ
0.287
2
ρ
0.273
***: significant at the 0.05 level

19
**: marginally significant at the 0.1 level
6. Conclusions
The most important finding was that 76.6% of the respondents reported choosing a specific
car intentionally. Among them, 69.7% stated that their motivation was to minimize the walking
distance to an exit when they disembarked at a destination station, 16.6% reported that they
sought to minimize the distance from the entrance when they boarded at an origin station, and
the remaining 13.5% reported that they wanted to pursue comfort while traveling.
The present study investigated the potential motivations behind these passenger preferences
for a specific car of a metro train. A two-level hierarchical decision structure was hypothesized to
account for passenger car choice. Four types of variables were selected to feed a nested logit
model based on the two-level hierarchy: individual-specific characteristics, trip-related variables,
physical environments, and psychometric propensities.
Regarding the upper-level decision, white collar young females with a compulsory trip
purpose were more likely to choose a specific car of a train in the morning peak hours. A
passenger’s awareness about the layout of his/her origin and destination stations proved to
expedite the intentional choice for a specific car of a train. For a lower-level decision, the location
of physical environments such as entrances/exits, escalators, and transfer gates proved to have a
considerable impact on the motivation for a passenger car choice based on minimizing the
walking distance. Mnemonic ability and prior travel experiences were found to influence the car
choice to minimize the walking distance at destination stations
The motivation to minimize the walking distance at destination stations was the most decisive
determinant of passenger choice for a specific car of a train. This implies that the even utilization
of train cars would be possible, at least for a projected metro line, if the precise forecast of travel
demand were incorporated with an optimal station layout considering passengers’ motivation to
minimize the walking distance at destination stations. Furthermore, there can be an operational
strategy to more evenly distribute the passenger load with no change in infrastructure. At least for
metro lines that have a platform longer than a running train, there might be the possibility to
accomplish a more evenly distributed passenger load by differentiating train-stop positions along
a platform for each station.
Controlling individual-specific characteristics and psychometric propensities could be another
key to a more even dispersion of passengers across individual cars. However, the present study
does not provide sufficient grounds to foment policy to force passengers to scatter evenly among
individual cars. Nevertheless, it might be possible to evenly distribute the passenger load without
changes in both the structure and operation of a metro line. The promotion of walking could be
introduced as an effective way to accomplish the load balance, particularly when associated with
health enhancement benefits. Further research should focus on excavating more latent variables to

20
accomplish an even utilization of specific cars of a train.
The uneven distribution of passengers across cars has been a well-known problem the world
over. As far as could be ascertained, however, the present study is the first to explore possible
solutions for the problem in a systematic manner at the individual passenger level. It is uncertain
that the findings of the present study are transferable to other contexts. The present study,
however, could trigger ensuing studies to find answers to the questions of transferability.

21
References
Ben-Akiva, M. and Lerman S., 1985. Discrete choice analysis, The MIT Press, Cambridge
Massachusetts
Ben-Akiva, M., Choudhury, C., Toledo, T., 2006. Modeling latent choices: application to driving
behavior. Paper presented at the 11th International Conference on Travel Behaviour Research,
Kyoto, Japan, 16–20
Ben-Akiva, M., Walker, J., Bernardino, A.T., Gopinath, D., Morikawa, T., Polydoropoulou, A., 2002.
Integration of choice and latent variable models. In: Mahmassani, H.S. (ed.) Perpetual Motion:
Travel Behaviour Research Opportunities and Application Challenges, pp. 431–470. Elsevier,
Amsterdam
Ben-Elia, E., Erev, I., Shiftan, Y., 2008. The combined effect of information and experience on drivers’
routechoice behavior. Transportation 35(2), 165–177
Currie, G., 2010. Quick and effective solution to rail overcrowding: free early bird ticket experience
in Melbourne, Australia. Transportation research record 2146(1): 35-42.
Hale, C. and P. Charles, 2009. Managing peak demand for passenger rail: A literature review. 32nd
Australasian Transport Research Forum (ATRF). Auckland, New Zealand, ATRF
Hirsch, L., Thompson, K., 2011. I can sit but I’d rather stand: Commuter’s experience of
crowdedness and fellow passenger behaviour in carriages on Australian metropolitan trains.
Proceedings of Australian Transport Research Forum 2011, Adelaide, Australia
Johansson, M.V., Heldt, T., Johansson, P., 2006. The effects of attitudes and personality trait on
mode choice. Transp. Res. Part A 40, 507–525
Jong, J.-C., Lai, Y.-C., Huang, S.-H., Chiang, P.-C., 2011. Development and Application of Rail Transit
Capacity Models in Taiwan. Transportation research record 2216, 125–138
Lee, I., Park, H., Sohn, K., 2012. Increasing the number of bicycle commuters. Proceedings of the
Institution of Civil Engineers - Transport 165, 63–72
Morikawa, T., Ben-Akiva, M., McFadden, D., 2002. Discrete choice models incorporating revealed
preferences and psychometric data. In: Franses, P.H., Montgomery, A.L. (eds.) Econometric Models

22
in Marketing, Advances in Econometrics, vol. 16, 29–55. Elsevier, Amsterdam
Otuzar, J. and Willumsen, L.G., 2004. Modelling Transport, Third Edition. John Wiley, Chichester, U.K.
Papinski, D., Scott, D.M., Doherty, S.T., 2009. Exploring the route choice decision-making process: a
comparison of pre-planned and observed routes obtained using person-based GPS. Transp. Res.
Part F 12(4), 347–358
Prato, C.G., Bekhor, S., Pronello, C., 2012. Latent variables and route choice behavior.
Transportation 39, 299-319
Pudney, P., Wardrop, A., 2008. Generating Train Plans with Problem Space Search, Lecture Notes in
Economics and Mathematical Systems 600 Part II, 195-207
Qi , Z., Baoming, H., Li, D., 2008. Modeling and simulation of passenger alighting and boarding
movement in Beijing metro stations. Transportation Research Part C 16, 635–649
Sohn, K., Yun, J., 2009. Separation of car-dependent commuters from normal-choice riders in
mode-choice analysis. Transportation 36, 423–436
Transit Co-operative Research Program (TCRP) Report 13, 1996. Rail transit capacity. Transportation
Research Board/National Academy Press
Transit Co-operative Research Program (TCRP) Report 100, 2003. Transit Capacity and Quality of
Service Manual. Transportation Research Board/National Academy Press
Vuchic, V. R., 1981. Urban Public Transportation Systems and Technology. Prentice Hall, Inc.,
Englewood Cliffs, N.J.